Update README.md
Browse files
README.md
CHANGED
@@ -39,11 +39,11 @@ Unfortunately, this does not work for dialog because conversational statements a
|
|
39 |
5. Diwank loves to eat Ishita's head.
|
40 |
|
41 |
**Dialog**:
|
42 |
-
> Diwank: Hey, what are we eating for dinner today?
|
43 |
-
> Ishita: Already? I thought we just ate lol
|
44 |
-
> Diwank: Yeah, some of us work hard and get hungy
|
45 |
-
> Ishita: Okay, what do you want to eat then?
|
46 |
-
> Diwank: I want to eat out but I am thinking of something light.
|
47 |
|
48 |
Now, a text/vector/hybrid search would probably match all 5 facts to this conversation but, as you can see, only facts 1 and 2 are relevant. The only way to get the correct fact, right now, is to ask an LLM like gpt-3.5 to "generate a query" for querying the database and then using that for similarity. Unfortunately, there are three big problems with that:
|
49 |
- It adds latency and cost.
|
@@ -60,6 +60,22 @@ This solves all of the three problems from the "query generation" method from ea
|
|
60 |
|
61 |
The "query generation" method is still far superior in quality but is too prohibitive (costly + slow) in normal circumstances and DFE solves that. :)
|
62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
## Usage (Sentence-Transformers)
|
64 |
|
65 |
Using this model becomes easy when you have [sentence-transformers](https://www.SBERT.net) installed:
|
@@ -72,13 +88,34 @@ Then you can use the model like this:
|
|
72 |
|
73 |
```python
|
74 |
from sentence_transformers import SentenceTransformer
|
75 |
-
sentences = ["This is an example sentence", "Each sentence is converted"]
|
76 |
|
77 |
-
|
78 |
-
|
79 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
```
|
81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
|
83 |
## Training
|
84 |
The model was trained with the parameters:
|
@@ -113,6 +150,7 @@ Parameters of the fit()-Method:
|
|
113 |
|
114 |
<!--- Describe how your model was evaluated -->
|
115 |
|
|
|
116 |
|
117 |
## Full Model Architecture
|
118 |
```
|
@@ -140,4 +178,6 @@ SentenceTransformer(
|
|
140 |
|
141 |
## Citing & Authors
|
142 |
|
143 |
-
|
|
|
|
|
|
39 |
5. Diwank loves to eat Ishita's head.
|
40 |
|
41 |
**Dialog**:
|
42 |
+
> Diwank: Hey, what are we eating for dinner today?
|
43 |
+
> Ishita: Already? I thought we just ate lol
|
44 |
+
> Diwank: Yeah, some of us work hard and get hungy
|
45 |
+
> Ishita: Okay, what do you want to eat then?
|
46 |
+
> Diwank: I want to eat out but I am thinking of something light.
|
47 |
|
48 |
Now, a text/vector/hybrid search would probably match all 5 facts to this conversation but, as you can see, only facts 1 and 2 are relevant. The only way to get the correct fact, right now, is to ask an LLM like gpt-3.5 to "generate a query" for querying the database and then using that for similarity. Unfortunately, there are three big problems with that:
|
49 |
- It adds latency and cost.
|
|
|
60 |
|
61 |
The "query generation" method is still far superior in quality but is too prohibitive (costly + slow) in normal circumstances and DFE solves that. :)
|
62 |
|
63 |
+
## Technical details
|
64 |
+
|
65 |
+
It inherits the base BERT model and pooling layer from BGE to generate 768-dimensional embeddings for input text.
|
66 |
+
|
67 |
+
DFE then adds an Asymmetric projection layer with separate dense layers for "dialog" and "fact" inputs:
|
68 |
+
|
69 |
+
Dialog inputs pass through 2x1536D tanh layers, a dropout layer, and another 1536D tanh layer before projecting back to 768 dimensions.
|
70 |
+
Fact inputs pass through similar 1536D tanh layers with dropout before projecting back to 768D.
|
71 |
+
This asymmetric architecture allows specialization of the embeddings for relevance matching between dialogs and facts.
|
72 |
+
|
73 |
+
DFE is trained with a triplet loss using the TripletDistanceMetric.EUCLIDEAN distance function and a margin of 5. It pulls dialog embeddings closer to positively matched fact embeddings, while pushing non-relevant pairs beyond the margin.
|
74 |
+
|
75 |
+
The model was trained for 12 epochs using the Lion optimizer with 100 warmup steps and a learning rate of 0.0001. No evaluation steps were used during training.
|
76 |
+
|
77 |
+
This approach teaches DFE to transform dialog and fact embeddings into a joint relevance space optimized for low-latency semantic matching. The specialized projections allow fast approximation of relevant facts for conversational dialog turns.
|
78 |
+
|
79 |
## Usage (Sentence-Transformers)
|
80 |
|
81 |
Using this model becomes easy when you have [sentence-transformers](https://www.SBERT.net) installed:
|
|
|
88 |
|
89 |
```python
|
90 |
from sentence_transformers import SentenceTransformer
|
|
|
91 |
|
92 |
+
dialog = """
|
93 |
+
Diwank: Hey, what are we eating for dinner today?
|
94 |
+
Ishita: Already? I thought we just ate lol
|
95 |
+
Diwank: Yeah, some of us work hard and get hungy
|
96 |
+
Ishita: Okay, what do you want to eat then?
|
97 |
+
Diwank: I want to eat out but I am thinking of something light.
|
98 |
+
""".strip()
|
99 |
+
|
100 |
+
facts = [
|
101 |
+
"Diwank likes Sushi.",
|
102 |
+
"Ishita does not like unnecessarily-pricey places restaurants",
|
103 |
+
"Diwank likes cooking.",
|
104 |
+
"Ishita is terrible at cooking.",
|
105 |
+
"Diwank loves to eat Ishita's head.",
|
106 |
+
]
|
107 |
+
|
108 |
+
model = SentenceTransformer("julep-ai/dfe-base-en")
|
109 |
+
dialog_embeddings = model.encode({"dialog": dialog})
|
110 |
+
fact_embeddings = model.encode([{"fact": fact} for fact in facts])
|
111 |
```
|
112 |
|
113 |
+
## Dataset
|
114 |
+
|
115 |
+
The model was trained on a custom dataset [julep-ai/dfe-stacked_samsum](https://huggingface.co/datasets/julep-ai/dfe-stacked_samsum) that we created from [stacked-summaries/stacked-samsum-1024](https://huggingface.co/datasets/stacked-summaries/stacked-samsum-1024) by:
|
116 |
+
1. Extracting summaries for corresponding dialogs to emulate "facts"
|
117 |
+
2. Then truncating the dialogs to emulate "missing information"
|
118 |
+
3. And then augmenting the dialogs using LLMs to emulate "additional information"
|
119 |
|
120 |
## Training
|
121 |
The model was trained with the parameters:
|
|
|
150 |
|
151 |
<!--- Describe how your model was evaluated -->
|
152 |
|
153 |
+
TBD
|
154 |
|
155 |
## Full Model Architecture
|
156 |
```
|
|
|
178 |
|
179 |
## Citing & Authors
|
180 |
|
181 |
+
```
|
182 |
+
Diwank Singh Tomer, Julep AI Inc. Dialog Fact Encoder (DFE). https://julep.ai (2023).
|
183 |
+
```
|